A customer writes: “The tracking page says delivered, but there’s nothing outside.” Before anyone drafts a reply, the message needs to reach the right queue.
Imagine three choices: delivery help, returns, or a person who can investigate. The first useful AI output could be a label pointing to one of those choices. A beautifully written apology would still leave the routing job unfinished.
What IBM’s latest release adds
On October 15, IBM announced Granite Switch 4.1, combining chat, decision-making and speech transcription in one model deployment. The published model card describes how to run it; this is software for developers to set up, rather than a new customer-service app you simply sign into.
One component is a decision adapter: a small trained addition to the underlying model. Given some information, a question and a fixed set of options, it scores the options. IBM’s adapter documentation describes reading those scores from the model’s next-token predictions, without generating a long response.
This is a new package for an established kind of task. Text classification already assigns labels to text—for example, identifying whether a review expresses positive or negative sentiment. The interesting design question is where a clear choice would help your workflow.
The menu matters
Return to our invented inbox. For the missing parcel, “delivery help” looks sensible. Now change the message to: “The parcel arrived damaged, and I’m leaving tomorrow. Can you replace it today?”
That request touches delivery, returns and an urgent exception. If the only options are delivery and returns, the system must pick from an incomplete menu. Adding “needs a person” gives it another possible route; it does not guarantee that the model will recognize every difficult case.
IBM’s documentation makes the underlying limitation explicit: the adapter selects among the supplied options, and decision quality depends on the information, question and option descriptions. Designing those choices is part of designing the system.
Also notice what a routing label does not settle. Sending a message to the returns queue says nothing about whether a refund is allowed. That requires the relevant order details and policy. In this example, routing and approving a refund are separate decisions.
A confidence score needs a reality check
Suppose a system assigns “delivery help” a score of 90%. You might read that as “this route will be right nine times out of ten.” That interpretation needs checking.
The scikit-learn calibration guide explains the connection: reliable probabilities should line up with observed outcomes across comparable predictions. If 100 messages receive scores around 90% for delivery, but only 60 actually belong there, those scores are overconfident. These numbers illustrate the idea; they are not measurements of Granite.
A useful trial would include straightforward requests, mixed requests and cases the available queues do not cover. Check the chosen route against a known answer, inspect mistakes by category, and measure the time to reach the correct team. A fast wrong turn can still delay the customer.
For our missing-parcel message, the goal is modest: get it to someone who can help. A short label can do that job. The explanation can come when the customer actually needs one.